Quick Answer
If you used ChatGPT, Claude, or Gemini to write an essay and are worried your teacher will know, this guide explains exactly what they look for. If you teach and want to recognize AI writing more reliably, the same 10 signals apply. Run any draft through our AI Detector to see the same signals scored automatically.
The 10 Telltale Signs
1. Uniform Sentence Length (Low Burstiness)
This is the single biggest signal. Human writers mix very short sentences with long winding ones. AI text clusters around 18 to 22 words per sentence with little variance. Read an AI paragraph aloud and you hear a steady metronome rhythm. Read a human paragraph and you hear a drumkit, with hits of different size.
Researchers call this burstiness. The DetectGPT paper out of Stanford (Mitchell et al. 2023) found that burstiness is the single strongest statistical signal in machine-generated text detection.
2. AI Cliche Phrases (delve into, tapestry of, in conclusion)
Large language models love a fixed vocabulary of filler. Teachers have memorized the list. The most flagged phrases:
- delve into (extremely overused by ChatGPT)
- tapestry of
- navigating the complexities
- in today's digital age
- robust framework
- leveraging
- ever-evolving
- comprehensive overview
- myriad of
- crucial role
- in conclusion (as the only closing transition)
One or two of these in a 500-word essay might pass. Five or more is a red flag.
3. Vague Examples Without Specifics
AI hallucinates generalities. A real student writes my history teacher mentioned the Treaty of Versailles last week. AI writes many historians believe the treaty had significant consequences. Specifics anchor human writing. Generic claims (studies show, many people believe, in recent times) without a named source, date, or example are an AI fingerprint.
4. Em-Dash and Semicolon Overuse
Punctuation is a fingerprint. AI overuses the em-dash character and the semicolon. Most undergraduate writers rarely reach for either. A 500-word essay with four or more em-dashes and two or more semicolons is statistically unusual for a typical student draft.
5. Predictable Structure (Intro + 3 Body + Conclusion)
AI defaults to the five-paragraph essay even when the prompt does not call for it. Three body paragraphs, each opening with a transition word like furthermore or moreover, each three to five sentences long, each with a topic sentence followed by two supporting sentences and one closing sentence. The symmetry is too clean.
6. Hollow Conclusions That Just Summarize the Intro
AI conclusions tend to restate what the essay already said. In conclusion, as discussed above, the topic is important for many reasons. A human conclusion usually adds something: a new angle, a personal stake, a question for the reader, or a concrete recommendation. An AI conclusion loops back without forward motion.
7. Generic Transitions (furthermore, moreover, additionally)
AI loves academic-sounding transition adverbs. Furthermore, moreover, additionally, in addition, in conclusion, as a result. Human writers usually mix these with sentence-break transitions (a period, then a new sentence) and varied openers. A draft where every paragraph begins with the same transition family signals a model wrote it.
8. Lack of Personal Voice or Opinion
Even in formal academic essays, human writers leave fingerprints: word choices, slight bias, a moment of emphasis. AI drafts read as if no one in particular wrote them. No I, no my, no clear stake. When a teacher reads two paragraphs and cannot guess what the writer believes, that is an AI tell.
9. Surface-Level Analysis Without Depth
AI summarizes well and analyzes poorly. It can describe what a poem says. It struggles with why a specific image choice matters in the context of the poet's biography. Teachers reading literature, history, or philosophy essays notice when the writing covers a topic broadly without the kind of close reading that proves a student engaged with the material.
10. Hallucinated Citations or Wrong Details
AI invents sources confidently. A made-up Smith and Jones 2019 study. A book that does not exist. A misattributed quote. Subject-matter teachers spot these instantly because they know the field. A student citing a fake paper is one of the clearest AI signals possible. Teachers do check citations.
Real Example: AI Essay vs Human Essay Side-by-Side
Same prompt: discuss the impact of social media on teenage mental health.
AI version (200 words):
In today's digital age, social media has become an integral part of teenage life, profoundly shaping how young individuals interact with the world around them. Numerous studies have shown that excessive use of platforms like Instagram, TikTok, and Snapchat can lead to a myriad of mental health challenges. Furthermore, the constant exposure to curated content often fosters feelings of inadequacy and comparison among adolescents. Moreover, the addictive nature of these platforms can result in reduced sleep quality, which in turn negatively impacts emotional well-being. It is also important to note that cyberbullying, a pervasive issue in the digital landscape, can have devastating consequences on teenage mental health. Additionally, the lack of meaningful face-to-face interactions may contribute to feelings of isolation and loneliness. While social media offers a tapestry of opportunities for connection and self-expression, navigating the complexities of its impact requires a balanced approach. Parents, educators, and policymakers must work together to foster healthier digital habits among teenagers. In conclusion, by leveraging awareness and implementing thoughtful strategies, society can mitigate the adverse effects of social media on adolescent mental health and empower young individuals to thrive in an ever-evolving digital world.
Human version (200 words):
My cousin is 14 and she checks Instagram before she gets out of bed. I asked her once how it makes her feel. She said tired. That stuck with me. The research backs her up. A 2023 study from Sapien Labs of 27,000 young adults found that the earlier someone got a smartphone, the worse their mental health was at age 18 to 24. The effect was sharpest for girls. Not all of social media is bad. My cousin found her closest friend through a fanfiction Discord, and they talk every day. But the part that scrolls past her face during dinner, the part she cannot stop checking, that is the part hurting her sleep. The fix is not to ban anything. It is to give teenagers a say in their own use, then back it up with parents who put their own phones away. The hardest classroom rule my school ever passed was no phones in the cafeteria. People hated it for a week. Then nobody mentioned it. Most kids I know, including my cousin, were quietly relieved.
What changed:the AI version has uniform 18 to 24 word sentences, six clear AI cliches (today's digital age, myriad of, navigating the complexities, tapestry of, leveraging, ever-evolving), no specific source, no personal stake, and a hollow conclusion. The human version swings from 4-word sentences (She said tired.) to 30-word sentences, cites one specific study with a number, names a specific platform use case, takes a position, and ends on a small concrete anecdote.
Why Teachers Can Tell So Quickly
Three reasons. First, pattern recognition. A teacher who grades 30 essays a week for ten years has read 15,000 essays. The AI rhythm reads differently from the bell curve of human student writing. Once you have seen the pattern, you cannot unsee it.
Second, vocabulary signatures. Researchers and English departments now circulate informal lists of AI cliche phrases, often around 28 words and phrases that show up disproportionately in ChatGPT output. When five or more appear in a single student essay, that is a statistical tell.
Third, subject knowledge. Teachers know their field. A made-up citation, a wrong date, or a quote that the named author never said tells a history professor exactly what happened. AI hallucinations are a gift to subject-matter graders.
What to Do If You Used AI
You have options, and they matter in this order.
- Read your school's AI policy first. Some schools allow AI use with disclosure. Some require it. Some ban it entirely. Disclosure is usually the safest path if you have not yet submitted.
- Rewrite into your own voice. Our full guide on how to humanize AI text walks through 9 specific techniques. The goal is not to fool detectors. It is writing that sounds like you.
- Add real specifics. Your own examples, a real source you read, a personal observation. Even one concrete detail per paragraph changes the read.
- Test with our AI Detector. Paste your revised draft into our AI Detector tool. Aim for a Likely Human or Uncertain verdict before you submit.
- If you already submitted and feel guilty, talk to your teacher. Most teachers respond better to a student who comes forward than to one who denies it under questioning.
What to Do If You Are a Teacher
Suspecting AI use is one thing. Acting on it fairly is another. Here is what experienced educators do.
- Use detectors as one signal, not as proof. A high score from any AI detector, including ours, is a reason to look closer. It is not by itself sufficient grounds for an academic integrity case.
- Read the obvious signs first. If the essay has six AI cliches, a hollow conclusion, and a hallucinated citation, those are stronger evidence than any score.
- Hold a conversation, not an interrogation. Ask the student to walk you through paragraph three. Ask where they found a specific claim. A student who wrote the essay can usually answer. A student who pasted it cannot.
- Know the false positive risk. Liang et al. (2023) found that GPT detectors flagged 61% of TOEFL essays by non-native English speakers as AI-generated. Students with formal writing styles, autism, or heavy Grammarly use are also vulnerable.
- Build a clear policy and share it on day one. Ambiguity creates more cheating than enforcement prevents.
The Quick Detection Checklist
Print this. Use it on any essay you suspect.
- Does sentence length vary wildly (5 to 40 words) or stay around 18 to 22?
- How many AI cliche phrases appear? (delve into, tapestry of, navigating the complexities)
- Are examples specific (named people, dates, sources) or vague (many believe, studies show)?
- How many em-dashes and semicolons in the first 500 words?
- Is the structure rigid five-paragraph or does it vary?
- Does the conclusion add anything or just restate the intro?
- Do you hear the student's voice or could anyone have written this?
- Does the analysis go below the surface or stay at description?
- Can you verify every cited source?
- What does our AI Detector say when you paste a sample paragraph?
Sources
- Mitchell, E., Lee, K., Khazatsky, A., Manning, C.D., & Finn, C. (2023). DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature. Stanford University.
- Gehrmann, S., Strobelt, H., & Rush, A.M. (2019). GLTR: Statistical Detection and Visualization of Generated Text. Harvard NLP / MIT-IBM Watson AI Lab.
- Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, Cell Press.
- Crothers, E., Japkowicz, N., & Viktor, H. (2023). Machine-Generated Text: A Comprehensive Survey of Threat Models and Detection Methods. ACM Computing Surveys.
- Stanford Institute for Human-Centered AI (2024). AI in Education: Policy and Practice Brief.